Minutes_2021_12_07

Numba Meeting: 2021-12-07

Attendees: Siu Kwan Lam, Graham Markall, Nick Riasanovsky, stuart, Todd A. Anderson, Val, Guilherme Leobas, Luk, brandon willard

NOTE: All communication is subject to the Numba Code of Conduct.

0. Feature Discussion

• PSA: numba devs Holiday hours

  • Next week is last meeting for 2021
  • After Dec 17 until Jan 11 of 2022, reduce support on Gitter, GH, Discourse...
  • Back to normal after Jan 11

• Getting 0.55.0 out the door this side of the holiday

  • #7560 refactor with detection for py3.10 merging soon
  • build/packaging problems: missing numpy packages for Python3.10
    • we will not ship py3.10 conda packages during the RC because of this
  • Pending PRs:
    • last of Py3.10
    • isinstance
    • and a few others if time allow

• Ask community about dropping py3.7 support.

  • per NEP 29, 0.55.0 can be the last version to support it.

• Last to discuss (10 min timebox): llvmlite to support multiple LLVM verions?

  • Support for Numba jit usecase, and non-Numba usecase.
  • Numba requires additional patches on stock LLVM
  • The plan:
    • make it easier to install llvmlite with user provided LLVM
    • maintainers will not make extra builds
  • Communicate to users what are supported use-cases and what aren't and how these will be supported in terms of priority
  • Potentially we could use conda-forge LLVM for testing "secondary" LLVMs e.g. in case of pull-requests for "secondary" LLVMs
  • If changes start being more on the C++ rather than on Python side, we may need to abandon ship?
  • Do we have an "exit strategy"? --> going back to "only a single version"
  • We will only accept secondary future versions, e.g. if we now support 11, we will accept 12 and 13

• Debugging LLVM optimization:

  • Gitter discussion: https://gitter.im/numba/numba?at=61ae7cca197fa95a1cacc28d
  • opt
    • conda install numba::llvmdev installs LLVM with its cmdline utils
    • NUMBA_OPT=0 turns optimization off for the process. func.inspect_llvm(signature) gets the LLVM IR for a function given its signature
    • Useful commands for seeing what LLVM optimization passes did: opt -O3 -print-after-all; opt -view-cfg

1. New Issues

• #7607 - CUDA: Linking PTX only works for eager compilation
• #7611 - njit with parallel=True causes fatal Python error and aborted
• #7612 - CUDA Benchmarking
• #7613 - CUDA typing mode
• #7614 - Make sure types.Type instances are not mutated
• #7615 - Significant slow down on njit'd functions when optional arguments not provided
• #7616 - Local variable gets overwritten causing wrong gdb debugging display
• #7620 - clamp tbb to 2021 on the 0.55 release branch

Closed Issues

• #7609 - Test suite on mainline@8be3456 is failing.

2. New PRs

• #7608 - [WIP] Updating TBB backend calls for tbb=2022.1
• #7617 - Numba gdb-python extension for printing
  • Graham will be testing this on CUDA
• #7619 - CUDA: Fix linking with PTX when compiling lazily
• #7621 - Add support for linking CUDA C / C++ with @cuda.jit kernels
  • Maybe even to look into NVRTC for internal use
• #7622 - Support fortran loop ordering for ufunc generation

Closed PRs

• #7610 - Fix #7609. Type should not be mutated.
• #7618 - Fix the doc build: docutils 0.18 not compatible with pinned sphinx

3. Next Release: Version 0.55.0/0.38.0, RC=Dec 1st

• Request for 0.55
  • broadcast_to https://github.com/numba/numba/pull/7119
    • driven by need for vectorize support in aesara
    • Now merged